[Ocfs2-devel] [nov 28] question of dlm_get_lock_resource()

[Sunil Mushran <sunil.mushran@oracle.com>]

On second thoughts, I should elaborate further. I'll do so sometime 
later today.

Sunil Mushran wrote:
> Read the purge_lockres code, the only place where we set this flag.
>
> Coly Li wrote:
>   
>> Sunil Mushran Wrote:
>>   
>>     
>>> No. dlm->spinlock prevents it racing the dlm_thread. Secondly, lookup
>>> also does a get.
>>>
>>> What scenario are you envisioning?
>>>     
>>>       
>> IMHO, locking dlm->spinlock does not mean protecting to tmpres. Even line 727 gets false, how can we
>> make sure in line 749 res->state is not modified to DLM_LOCK_RES_DROPPING_REF by other process ?
>>
>> I can not find the magic which make everything work in order, something important must be missed by me.
>>
>> Thanks.
>>
>>   
>>     
>>> Coly Li wrote:
>>>     
>>>       
>>>> Hi list,
>>>>
>>>> When I read code of dlm_get_lock_resource(), there is something not
>>>> clear to me.
>>>>
>>>>  719 lookup:
>>>>  720         spin_lock(&dlm->spinlock);
>>>>  721         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen,
>>>> hash);
>>>>  722         if (tmpres) {
>>>>  723                 int dropping_ref = 0;
>>>>  724
>>>>  725                 spin_lock(&tmpres->spinlock);
>>>>  726                 if (tmpres->owner == dlm->node_num) {
>>>>  727                         BUG_ON(tmpres->state &
>>>> DLM_LOCK_RES_DROPPING_REF);
>>>>  728                         dlm_lockres_grab_inflight_ref(dlm, tmpres);
>>>>  729                 } else if (tmpres->state &
>>>> DLM_LOCK_RES_DROPPING_REF)
>>>>  730                         dropping_ref = 1;
>>>>  731                 spin_unlock(&tmpres->spinlock);
>>>>  732                 spin_unlock(&dlm->spinlock);
>>>>  733
>>>>  734                 /* wait until done messaging the master, drop our
>>>> ref to allow
>>>>  735                  * the lockres to be purged, start over. */
>>>>  736                 if (dropping_ref) {
>>>>  737                         spin_lock(&tmpres->spinlock);
>>>>  738                         __dlm_wait_on_lockres_flags(tmpres,
>>>> DLM_LOCK_RES_DROPPING_REF);
>>>>  739                         spin_unlock(&tmpres->spinlock);
>>>>  740                         dlm_lockres_put(tmpres);
>>>>  741                         tmpres = NULL;
>>>>  742                         goto lookup;
>>>>  743                 }
>>>>  744
>>>>  745                 mlog(0, "found in hash!\n");
>>>>  746                 if (res)
>>>>  747                         dlm_lockres_put(res);
>>>>  748                 res = tmpres;
>>>>  749                 goto leave;
>>>>  750         }
>>>>
>>>> If at line 721 tmpres found from hash, and its state is not
>>>> DLM_LOCK_RES_DROPPING_REF (dropping_ref
>>>> is 0), between line 733 and 748, is it possible to set tmpres->state
>>>> to DLM_LOCK_RES_DROPPING_REF ?
>>>>
>>>> I don't see any protection for this race, maybe there is something I
>>>> missed. Can anybody give me a
>>>> hint ?
>>>>       
>>>>         
>>   
>>     
>
>
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> http://oss.oracle.com/mailman/listinfo/ocfs2-devel
>